Find the people to contact at every company

Export everyone at each company, then choose. Filtering the export loses the people who matter: seniority is guessed from the job title and is missing or wrong for about 1 in 8 decision makers (roughly 88% accurate), title filters match whole words only, and most people have no email in the database until they are enriched. lookup_decision_makers returns only people with a verified email, so it is not used here.

Keep the user with you. Before each stage, say in one line what it does; after it, show the result as a table: people pulled and dropped per company, then the picks with their titles, then emails found.

1. Brief

Take the company list: keep.csv from getleads-company-list (https://getleads.io/company-list/) (it carries company_id, which makes step 3 exact), or the user's own file with a website or domain per company.

Ask what they sell and who signs off on it at their customers, then write brief.json:

{
  "offer": "One or two sentences: what the user sells.",
  "buyer": "Who signs off on it at their customers, by company size if that changes it.",
  "avoid": ["Optional: people never to contact, one condition per item, e.g. an intern"]
}

The titles of the people who actually bought at their existing customers make the best buyer.

2. Pull everyone

One export_contacts per batch of companies, with domains set to each company's website host and no seniority, title, email or max_per_company filter, and these columns:

contact_id, full_name, current_title, seniority_level, department, linkedin_headline, about_person, linkedin_url, work_email, email_status, role_start_date, company_id, company_name, company_website, company_linkedin_url, linkedin_employee_count_exact

• About 25 companies per batch, fewer once their headcounts add up past 5,000. Start every batch's export, then check them, so they run together. Save each as pulls/<batch>.csv.
• Don't count first: people-at-company counts often time out.
• A company whose website is a page on another company's site (airbus.com/…), a link shortener or a site builder would pull that site's whole staff by domain. Pull those by company_name instead; step 3 drops namesakes.
• An export that returns exactly 50,000 rows is full, whatever cap_reason says: split the batch.

3. Org charts

python3 scripts/org_charts.py --companies keep.csv --pulls pulls/*.csv --out run

It keeps only people who work at each company, matched on company_id when the list has it, otherwise the company's website, then its name. People exported under a company's domain who work elsewhere or nowhere ("Retired") are dropped. Companies of 50 or fewer keep everyone. Bigger ones keep the leadership layer, up to 60 people, most senior first: owners, founders, C-level, VPs, heads and directors, plus anyone with no seniority recorded whose title or headline reads senior.

Show the user summary.json: rows pulled, dropped as another company, and companies with nobody.

4. Pick

It needs pip install typesafe-sdk and TYPESAFE_API_KEY (typesafe.ai) in the environment, .env or ~/.agents/.env.

python3 scripts/pick_contacts.py --charts run/org_charts.jsonl --brief brief.json --out run --limit 10

Jev rates each person in a company's org chart as having no impact on buying the offer, having an impact, or deciding, reading them against the rest of the chart and the company's size: at a small company more of the team weighs in, at a larger one mainly the team the offer serves. By default everyone likely to have an impact is kept, deciders first. --mode top keeps one decider plus up to two backups instead (--picks sets how many); use it when the user wants one contact per company.

Read the ten companies' picks with the user against their charts. If the picks lean the wrong way, fix buyer and re-run (answers are cached); then drop --limit. 25 companies take seconds. Re-run until failed_requests and not_fully_judged in summary.json are 0: a company is never picked from a partly judged chart.

No TypeSafe key: --no-judge writes run/to_judge.jsonl. Judge each company's chart yourself against the brief, keeping who each line's keep says (it follows --mode), write picks.json as {"<key>": ["p3", "p0"]} (deciders first, [] when nobody fits), then run the same command with --agent-picks picks.json instead of --no-judge to write the outputs. role and the scores stay blank for your picks.

5. Emails and delivery

Most picks have no email in the export. Run getleads_get_emails_from_linkedin_batch on the picks' linkedin_urls. The answer also says whether each person still holds the role: drop a pick whose current_role_status is past.

Deliver from run/:

• contacts.csv: the people picked at each company, deciders first, with their role (decides or has an impact) and the emails found.
• no_pick.csv: companies with no pick, the reason (nobody in the data owns the area, or the best person was ruled out by an avoid rule) and the best person found. Offer those as they are or leave them out.